Abstract
Multiple instance learning (MIL) has been widely used for representingwhole-slide pathology images. However, spatial, semantic, and decisionentanglements among instances limit its representation and interpretability. Toaddress these challenges, we propose a latent factor grouping-boostedcluster-reasoning instance disentangled learning framework for whole-slideimage (WSI) interpretable representation in three phases. First, we introduce anovel positive semi-definite latent factor grouping that maps instances into alatent subspace, effectively mitigating spatial entanglement in MIL. Toalleviate semantic entanglement, we employs instance probability counterfactualinference and optimization via cluster-reasoning instance disentangling.Finally, we employ a generalized linear weighted decision via instance effectre-weighting to address decision entanglement. Extensive experiments onmulticentre datasets demonstrate that our model outperforms allstate-of-the-art models. Moreover, it attains pathologist-alignedinterpretability through disentangled representations and a transparentdecision-making process.